Formats¶

convert picks the format from the file's first bytes and its extension. For bytes, you name it: format="docx".

Documents¶

Format	Extensions	Out
Word	`.docx`	Headings, paragraphs, bold/italic, lists, tables, links, images, comments
PowerPoint	`.pptx`	One `## Slide N` per slide, tables, speaker notes, images, comments
Excel	`.xlsx` `.xls`	One `## Sheet` per sheet, as a table
PDF	`.pdf`	Every page, in reading order. Not native — needs `anytomd[pdf]`

Web and data¶

Format	Extensions	Out
HTML	`.html` `.htm`	Headings, tables, lists, links, code blocks; `<title>` becomes `document.title`
CSV	`.csv`	One table
Jupyter	`.ipynb`	Markdown cells as-is, code cells fenced
JSON	`.json`	Pretty-printed, fenced
XML	`.xml`	Pretty-printed, fenced

Text and code¶

Format	Extensions	Out
Text	`.txt` `.md` `.rst` `.log` `.yaml` `.toml` `.ini` …	As-is; UTF-8, UTF-16 and Windows-1252 detected
Code	`.py` `.rs` `.js` `.ts` `.go` `.java` `.c` …	Fenced, with its language

Images¶

.png .jpg .gif .webp .bmp .tiff .svg .heic .avif, alone or inside a DOCX, PPTX or XLSX.

Installed	An image with text	An image without text
nothing	`![](image.png)`	`![](image.png)`
`[ocr]`	its text	`""`
`[ocr]` + `[vlm]`	its text	`Image description: A red square on white.`
`[vlm]` only	`Image description: …`	`Image description: …`

OCR always goes first: it is local, free and exact. The vision model is asked only when OCR finds no letters, and its answer is always labeled Image description:, so you can tell it from text that was really in the image.

import anytomd as ad

ad.convert("invoice.png").markdown   # e.g. 'Invoice 2026\nAmount due: 1.240,00 EUR'
ad.convert("cat.png").markdown       # e.g. 'Image description: A tabby cat asleep on a wooden chair.'

Inside documents, the same rule fills each image's alt text, on one line:

![Invoice 2026 Amount due: 1.240,00 EUR](image1.png)
![Image description: A tabby cat asleep on a wooden chair.](image2.png)

When neither finds anything, the author's own alt text stays: ![Sales chart](chart.png).

Vision needs a model → VisionSpec. Turn OCR off with ocr=False.

PDF¶

The engine leaves PDF out on purpose: LLMs read it natively. To convert it here:

pip install "anytomd[pdf]"

document = ad.convert("report.pdf")
document.markdown   # every page, joined
document.pages      # 12

Text layers are read directly; scanned pages go through OCR (ocr=False skips it). Without the extra:

ad.convert("report.pdf")
# MissingExtraError: PDF needs the 'pdf' extra: pip install 'anytomd[pdf]'

Text recovery¶

Every supported format is generated from one and the same 103-word passage, then converted back and scored on how many of those words come out. It is a comparison between formats, not an absolute benchmark: the passage is fixed and committed, so the bars show where each format's text lands relative to the others.

Text recovered per format

Text and structured formats recover the passage in full. The gap is the OCR row: a rendered image is read back through the OCR engine, and some words survive imperfectly. The SVG row is zero — the engine treats an image as pixels and does not read vector <text>, so nothing comes back until the binding can rasterize it.

The measurement lives in tools/fidelity.py; docs/fidelity.json is the committed baseline. Re-run and compare with:

make fidelity            # re-measure, fail on any regression against the baseline
make fidelity ARGS=--update   # rewrite the baseline and the chart

Not supported¶

Plain ZIP archives, EPUB, binaries → UnknownFormatError.